How to Track Meals You Didn't Cook: Restaurants and Takeaway

The meals that break a food log

A food log is easy for food you cooked yourself. You know what went in the pan, the packet has numbers printed on the side, and the scale is on the counter. Eat somewhere else and all three of those disappear at once.

This is not an edge case. The 2011 study that put restaurant food in a bomb calorimeter opens by noting that restaurant food supplies roughly 35% of daily energy intake in the United States.² A Canadian national survey of 20,475 people found 21.8% had eaten food from a restaurant on the previous day, rising to 27.7% among men aged 19 to 54.⁷ For a lot of people that is one day in four or five containing a meal nobody has any direct information about.

What that does to a log has been measured. Researchers waited outside 89 fast food restaurants across four New England cities and asked 1,877 adults leaving with dinner to estimate the calories in what they were carrying. The meals averaged 836 calories. The estimates were low by an average of 175 (95% confidence interval 145 to 205), and the 1,178 adolescents surveyed were low by 259.¹

The number to carry around is not the 175. It is that underestimation increased substantially as the actual calorie content of the meal increased.¹ The gap grows with the plate, so it opens widest on exactly the meals that move a weekly total.

The limits matter. That study is cross-sectional, covers six chains in one region of the United States, and takes its true calorie figures from the chains' own published data rather than from measuring the food. So it gives you a direction and a rough size, not an account of what happened to you on Thursday.

What a menu calorie count is worth

Where a number is printed, it holds up better than you might expect. What it does badly, it does badly in one particular place.

Researchers bought 269 items from 42 randomly selected quick serve and sit down restaurants in Massachusetts, Arkansas and Indiana, and burned them. Averaged across everything, stated and measured energy were not significantly different: a gap of 10 kcal per portion, with a confidence interval running from 15 below to 34 above.² So restaurants as a group are not inventing their numbers.

Underneath that average, 50 of the 269 items, 19% of them, measured at least 100 kcal per portion above what the menu claimed. The researchers went back for a second sample of the worst offenders and got the same answer: those items ran 289 kcal over on the first pass and 258 kcal over on the second.² It was the dish, not a bad night in the kitchen.

The error also has a shape. Foods with lower stated energy contents measured higher than stated, while foods with higher stated contents measured lower.² The item advertised as the light choice is the one most likely to be under-labelled, which is the worst place that bias could possibly sit.

An earlier study from the same lab went at diet food directly. Twenty-nine reduced-energy restaurant items measured 18% above their stated values, and ten supermarket frozen meals 8% above, though the authors are careful to say the differences did not reach statistical significance because the variation between items was so large.³ The figure from that paper I have not managed to forget is a different one: once the free side dishes that came with the entrees were counted, the food actually served averaged 245% of the stated value for the entree alone.³

So a posted number is a fair starting point for a main dish, a poor one for anything sold as light, and it describes the item you ordered rather than the plate that arrives.

When there is no number at all

In the US, federal menu labelling rules cover chains with 20 or more locations.⁸ Independent and small chain restaurants, which account for roughly half of all restaurant locations, are not covered.⁴

Those got measured too. The 42 most frequently purchased meals across nine restaurant categories, 157 individual meals in total, bought from randomly selected independent and small chain restaurants within 15 miles of downtown Boston, and burned. The mean meal came to 1,327 kcal (95% CI 1,248 to 1,406), which the authors put at 66% of a typical day's energy requirement. Some 7.6% of meals exceeded a full day's requirement on their own.⁴

Two details there matter more than the average. The same meal from the same restaurant varied with an average standard deviation of 271 kcal, so the portion is not a fixed quantity you could learn by repetition.⁴ And in the subset the researchers could match directly, the real meals contained 19% more energy than national food database entries for the equivalent dishes.⁴ Looking the dish up in a database is not a workaround for the missing menu number. It is a second guess, biased in the same direction as the first.

One American city in 2013 and 157 meals is not the world, and I would not transplant those absolute figures to a curry house in Leeds. But bomb calorimetry is about as close to ground truth as food data gets, and where this group found error, it ran the same way each time: more energy on the plate than in the number attached to it. If you want the wider picture of where tracking error comes from, I went through all of it in how accurate calorie tracking apps actually are.

How well can anyone judge a portion?

With no menu number, and a database that runs low on this exact category of food, you are estimating. Two studies give a sense of how that goes, and neither is flattering.

In Tokyo, 54 adults aged 18 to 33 were served fourteen foods in the amounts they usually ate. A researcher weighed each portion once they had gone. The next day, the participants estimated those amounts from a set of food atlas photographs. The average relative difference came out at 8.8%, which sounds reassuring until you look at the spread: 51.6% of estimates landed within 25% of the true weight and 81.9% within 50%. Individual foods ran from a 29.8% underestimate for curry sauce to a 34.0% overestimate for margarine, and agreement got worse as servings got larger.⁵

That is close to a best case. These were portions sized to what the participants normally eat and seen the day before, judged with reference photographs in front of them. Half the estimates were still more than a quarter out.

The second study is about the photograph rather than the person. Eighty-two Korean adults looked at a laid out meal for three minutes, walked to another room, and picked the photo matching the amount they had seen. Each food had been shot from three angles. Cooked rice was judged correctly 74.4% of the time at 45 degrees, rising to 85.4% when several angles were available. Vegetables reached 53.7% with the angles combined. Soup was worst at every angle and was overestimated more often than the rest. Drinks did best at 70 degrees, the widest angle they tried.⁶

Small samples, Japanese and Korean foods, controlled reference images: do not read those percentages as universal. What survives the caveats is the ranking. Discrete solid food is judgeable, a bowl of liquid is not, and the angle a plate is seen from changes the answer. The read across to photographing your own dinner is mine rather than the authors', who were validating survey instruments.

How to track meals you didn't cook

None of that is an argument for giving up on the log. It is an argument for spending the effort somewhere else. Six rules, roughly in the order they come up.

• Log it while the food is still in front of you. The estimates in the New England study were made by people holding the bag, and they were still 175 calories low.¹ Nothing about the meal gets clearer by bedtime.
• Use the posted number when there is one, and treat the light options with suspicion. That is where the measured under-labelling concentrates.²
• Count the sides, the sauce and the drink as separate items. Sides produced the single largest number in this literature: they took the energy actually served to 245% of the entree's stated value.³
• Where nothing is posted, do not go hunting for the perfect database entry. For restaurant food those entries measured about 19% low.⁴ Pick a close match, round it up, move on.
• Photograph the plate as served, before you touch it, at an angle rather than straight down, with a fork or a hand in shot for scale. Everything after the first bite is a reconstruction.
• Round up when you are unsure. Most of the errors above run one way: the diners underestimated, the light items were understated, and the database entries came in low. The exception is that items with high stated calories tended to measure below their labels,² so the very largest posted numbers may already carry some slack.

What a meal tracker app can do here, and what it cannot

I build a calorie app, so read this as interested rather than neutral. What I can usefully offer is a clear account of which half of the problem software actually touches.

A search box is at its worst exactly where you need it most. It works well on a packet, because a packet is a named product with a fixed recipe, and it degrades on a plate assembled by a stranger. That is the argument for photographing the meal instead, and it is why MyPlate has no food database and no barcode scanner: you point the camera at the plate and a model estimates calories, protein, carbohydrate and fat from the picture. There is nothing to search for, which is the whole advantage for a restaurant plate and a straight disadvantage for a tin of beans with the numbers printed on the label.

What a photograph cannot see is the part that would help most. The oil in the pan and the butter finishing the steak are gone by the time the plate reaches you, and they are a large part of why the restaurant version of a dish outweighs the one you make at home. Of the four numbers a scan gives back, fat is the one I would trust least, which is also where those hidden calories land.

What an app can do is make the correction cheap. A scan in MyPlate comes back as a list of ingredients with weights, and every weight and every macro on that list can be edited; change a weight and that row's macros rescale at its own per 100 g rate. For a plate you did not cook, that is usually the only edit worth making, because the model has generally identified the food and guessed the quantity. If you got home before you remembered to log it, the date on a saved meal is editable too, so a late entry still lands on the day you ate it. That is the same argument I made about how much time these apps ask for: a correction that takes five seconds is what lets an estimate be good enough.

One last thing costs nothing at all. Using nationally representative US data from 2007 to 2018, people who reported using menu labels reported eating 202 fewer calories at fast food restaurants and 181 fewer at sit-down restaurants than those who did not.⁸ That is an observational association with self-reported intake on both sides, so it cannot separate reading the label from being the sort of person who reads labels, and self-reported intake runs low in general. It still points at the cheapest thing available: read the number before you order rather than after.

Common questions

Log the meal, not the recipe

The meals you did not cook are where a food log stops being an approximation and starts being a guess, and no amount of care on your part fixes that. Nobody weighed anything in the kitchen, and the menu number, where one exists, describes a specification rather than the plate in front of you. Even your own eye, holding a reference photograph of a portion you saw the day before, lands within a quarter of the real amount about half the time.⁵

So put the effort where it pays. Log the meal at all, log the sides with it, log it before you eat rather than after, and let the number be generous. A week of approximate entries tells you something real about how much you eat. Six precise days and a blank Saturday do not.

One caveat matters more on this topic than on most. If tracking a meal out has started turning dinner with other people into something anxious, put the app away for the evening. If that feeling keeps coming back, talk to your doctor or a registered dietitian, because at that point the problem is not which number you wrote down.

Sources

Every figure above traces to one of these. If you find a number that doesn't match the source it claims, tell me and I'll correct it.

1. Block JP, Condon SK, Kleinman K, et al. (2013). Consumers' estimation of calorie content at fast food restaurants: cross sectional observational study. BMJ, 346:f2907. Read on PubMed →
2. Urban LE, McCrory MA, Dallal GE, et al. (2011). Accuracy of stated energy contents of restaurant foods. JAMA, 306(3):287-293. Read on PubMed →
3. Urban LE, Dallal GE, Robinson LM, et al. (2010). The accuracy of stated energy contents of reduced-energy, commercially prepared foods. Journal of the American Dietetic Association, 110(1):116-123. Read on PubMed →
4. Urban LE, Lichtenstein AH, Gary CE, et al. (2013). The energy content of restaurant foods without stated calorie information. JAMA Internal Medicine, 173(14):1292-1299. Read on PubMed →
5. Shinozaki N, Murakami K (2022). Accuracy of estimates of serving size using digitally displayed food photographs among Japanese adults. Journal of Nutritional Science, 11:e105. Read on PubMed →
6. Choi IY, Kim MH (2025). Evaluating food portion estimation accuracy with multi-angle photographs. Nutrition Research and Practice, 19(4):605-620. Read on PubMed →
7. Polsky JY, Garriguet D (2021). Eating away from home in Canada: impact on dietary intake. Health Reports, 32(8):18-26. Read on PubMed →
8. Joshi R (2024). Prevalence of menu label use and its association with calorie intake among US adults. Appetite, 200:107577. Read on PubMed →

The formulas and limits behind MyPlate's own numbers, with their citations, are on the health sources page.